Semantic search has developed quickly as the need for accurate information retrieval has increased in a variety of fields, from expert knowledge systems to web search engines. Conventional search methods that rely solely on keywords frequently fail to understand user intent and contextual hints. This survey focuses on recent advances in Transformer-based models, such as BERT, RoBERTa, T5, and GPT, which leverage self-attention mechanisms and contextual embeddings to deliver heightened precision and recall across diverse domains. Key architectural elements underlying these models are discussed, including dual-encoder and cross-encoder frameworks, and how Dense Passage Retrieval extends their capabilities to large-scale applications is examined. Practical considerations, such as domain adaptation and fine-tuning strategies, are reviewed to highlight their impact on real-world deployment. Benchmark evaluations (e.g., MS MARCO, TREC, and BEIR) are also presented to illustrate performance gains over traditional Information Retrieval methods and explore ongoing challenges involving interpretability, bias, and resource-intensive training. Lastly, emerging trends—multimodal semantic search, personalized retrieval, and continual learning—that promise to shape the future of AI-driven information retrieval are identified for more efficient and interpretable semantic search.